Creating Synthetic Datasets for Deep Learning used in Machine Vision

Abstract

Deep Learning, a subfield of Artificial Intelligence, has enabled a wide range of applications in machine vision, including manufacturing quality control. A critical aspect of Machine Vision using Deep Learning algorithms is the availability of large amounts of visual training data. Collecting training data is time consuming, especially in cases where visual training data is difficult to collect. For example, machine vision systems for object detection in quality inspection need massive datasets with various light, rotation and viewpoints parameters to recognize components. Only the training phases of data collection and annotation are responsible for about 80 % of the time spent. Through advances in synthetically generated training data, the lack of available training data is alleviated and effort for data collection and annotation is reduced. This paper presents an innovative approach to generate synthetic training data using CAD-Tools and rendering software. In this work, synthetic training data is generated using the approach of domain randomization. Domain randomization introduces diversity and variability in training dataset, which has proven to improve the robustness and generalization of deep learning models. The approach is evaluated in a case example. Images of an assembly step are classified in terms of correct or incorrect assembly. The results show that synthetically generated training data used to train deep learning algorithms are fundamentally suitable for Machine Vision. Within this study deep learning models trained with synthetic image reached equally good results as models trained using real-world data. Variation using generated non-realistic image data is able to force neuronal nets to learn features of the object of interest. Synthetically generated training data therefore is an inexpensive source of data avoiding the need to collect and annotate large amounts of training data. Overall, the integration of synthetically generated image data has paved the way for new possibilities in visual application domains.